AI-driven framework to reconstruct real-time 3-D geological models for In-Situ Exploration of Critical Raw Materials

With a focus on geo-modeling applications for sustainable deep mineral exploration, we propose affordable, but still accurate subsurface modeling technique that can generates realistic 3-D geological models. Conventional geostatistical methods based on two-point statistics, often compromise their performance in deep and structurally complex geological settings mostly due to limitation in modelling complex spatial continuity patterns. On the other hand, deep generative modelling techniques, such as generative adversarial networks (GAN), allow to predict complex spatial patterns but have difficulties to create large-scales models in three-dimensions and be locally conditioned by observations.

We introduce a deep generative framework that adapts conditional GANs with spatially adaptive normalization (cGAN–SPADE) for 3-D geological modeling under sparse and evolving data conditions to predict high resolution subsurface models with real-time data assimilation capabilities. The goal is to generate geo-models based on a priori geological information (i.e., expected geometries and probability maps) with real-time model update as new data are acquired during drilling.

The cGAN-SAPDE is trained with samples based on prior geological knowledge and existing borehole experimental data. Training proceeds through a generator and discriminator scheme in which generator produces new models based on input training data while the discriminator output is the probability of input image being real based on the corresponding conditioning map.

A conditioning map is introduced at each generator’s layer, where it modulates the intermediate activations using SPADE normalization. This mechanism injects spatially varying conditioning information into the network, enabling the generator to preserve structural coherence and fine-grained spatial details in the synthesized outputs.

Experimental results on industry-standard challenging 3-D synthetic data sets show the ability of the network to predict high-resolution 3-D geological models that simultaneously match a priori information and direct measurements acquired in real-time scenario.

This project has received funding from the European Union’s Horizon Europe Research and Innovation Program under the Grant Agreement No.101178775